Autonomous-engineer products plan issues, edit repos, and open pull requests with variable human oversight.
After 2024 demos, 2025 saw more stable software agents: Devin-like systems, Codex agents, and rivals wired into CI, tickets, and internal browsers.
Real metrics—merged PR rate, time-to-green—replaced hype. Enterprises defined sandboxes and fine permissions.
The human developer did not vanish; they shifted to architecture, review, and designing tools for agents.
World-model research sought dynamics prediction for planning—beyond pretty pixel generation.
Learned emulators accelerated climate and materials simulations, complementing classic HPC.
New Llama generations improved multimodality and agents while keeping downloadable weights as industrial strategy.